AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians

TL;DR

AtlasGS tackles the challenge of globally consistent 3D reconstruction in indoor and urban scenes with low-texture regions. It introduces an Atlanta-world guided implicit-structured Gaussian Splatting approach that decodes Gaussians from a sparse feature grid, with Gaussian semantic lifting to predict floor/ceiling/wall semantics. The method couples 3D global planar regularization and 2D local surface regularization with explicit plane indicators to enforce consistent geometry across views, while preserving high-frequency details and rendering efficiency. Experiments on indoor (ScanNet, Replica, ScanNet++) and outdoor (MatrixCity) datasets show state-of-the-art surface reconstruction and novel view synthesis, outperforming both neural implicit and Gaussian-based baselines. This approach advances robust digital twins and robotics by enabling accurate, textureless-region geometry in complex environments.

Abstract

3D reconstruction of indoor and urban environments is a prominent research topic with various downstream applications. However, existing geometric priors for addressing low-texture regions in indoor and urban settings often lack global consistency. Moreover, Gaussian Splatting and implicit SDF fields often suffer from discontinuities or exhibit computational inefficiencies, resulting in a loss of detail. To address these issues, we propose an Atlanta-world guided implicit-structured Gaussian Splatting that achieves smooth indoor and urban scene reconstruction while preserving high-frequency details and rendering efficiency. By leveraging the Atlanta-world model, we ensure the accurate surface reconstruction for low-texture regions, while the proposed novel implicit-structured GS representations provide smoothness without sacrificing efficiency and high-frequency details. Specifically, we propose a semantic GS representation to predict the probability of all semantic regions and deploy a structure plane regularization with learnable plane indicators for global accurate surface reconstruction. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in both indoor and urban scenes, delivering superior surface reconstruction quality.

Figures (11)

• Figure 1: Qualitative comparison of indoor and outdoor reconstruction. For each scene, we visualize both the reconstructed geometry and the surface normals. As highlighted by the orange boxes, our method generates significantly smoother surfaces while capturing finer geometric details, clearly outperforming the compared methods.
• Figure 2: Overview of AtlasGS. Given posed images and SfM points, we construct a sparse feature grid and represent scenes as implicit-structured Gaussians. Attributes are decoded using attribute decoder and semantic decoder, followed by rasterization and supervision with RGB images, monocular geometry priors, and semantic maps. To address multiview inconsistency in textureless regions, we introduce learnable explicit plane indicators based on the Atlanta world assumption schindler2004atlanta. The indicators refine the global scene structure by regularizing Gaussian positions and orientations using 3D global planar and 2D local surface losses, ensuring alignment with structural elements such as walls, floors, and ceilings.
• Figure 3: Qualitative comparison of indoor scene reconstruction. We show the reconstruction performance of the baselines and our approach on ScanNet dai:2017:scannet, ScanNet++ scanet++, and Replica julian:2019:replica datasets. As highlighted in the boxes, our approach maintains local smoothness and preserves high frequency. The red dashed boxes mark regions that are zoomed in below for closer inspection.
• Figure 4: Qualitative comparison of outdoor scene reconstruction. As highlighted in the boxes, our approach can produce detailed and noise-free surfaces in textureless regions.
• Figure 5: Qualitative comparison of novel view synthesis. We show the novel view synthesis results of different Gaussian splatting-based approaches on ScanNet++ scanet++ and Replica julian:2019:replica datasets. Our method can obtain higher-fidelity rendering results with less noisy information than the baselines.